Commit 3d680f2e authored by Sebastian Nickels's avatar Sebastian Nickels

Merge

parents b85b4bce d3c1bc00
......@@ -8,7 +8,7 @@
<groupId>de.monticore.lang.monticar</groupId>
<artifactId>embedded-montiarc-emadl-generator</artifactId>
<version>0.3.3-SNAPSHOT</version>
<version>0.3.4-SNAPSHOT</version>
<!-- == PROJECT DEPENDENCIES ============================================= -->
......
......@@ -9,6 +9,7 @@ import de.monticore.lang.monticar.cnnarch.mxnetgenerator.CNNArch2MxNet;
import de.monticore.lang.monticar.cnnarch.caffe2generator.CNNArch2Caffe2;
import de.monticore.lang.monticar.cnnarch.mxnetgenerator.CNNTrain2MxNet;
import de.monticore.lang.monticar.cnnarch.caffe2generator.CNNTrain2Caffe2;
import de.monticore.lang.monticar.emadl.generator.reinforcementlearning.RewardFunctionCppGenerator;
import java.util.Optional;
......
......@@ -34,6 +34,8 @@ import de.monticore.lang.monticar.cnnarch._symboltable.ArchitectureSymbol;
import de.monticore.lang.monticar.cnnarch._symboltable.SerialCompositeElementSymbol;
import de.monticore.lang.monticar.cnnarch.gluongenerator.CNNTrain2Gluon;
import de.monticore.lang.monticar.cnnarch.gluongenerator.annotations.ArchitectureAdapter;
import de.monticore.lang.monticar.cnntrain._cocos.CNNTrainCoCoChecker;
import de.monticore.lang.monticar.cnntrain._cocos.CNNTrainCocos;
import de.monticore.lang.monticar.cnntrain._symboltable.ConfigurationSymbol;
import de.monticore.lang.monticar.emadl._cocos.EMADLCocos;
import de.monticore.lang.monticar.generator.FileContent;
......@@ -115,6 +117,18 @@ public class EMADLGenerator {
processedArchitecture = new HashMap<>();
setModelsPath( modelPath );
TaggingResolver symtab = EMADLAbstractSymtab.createSymTabAndTaggingResolver(getModelsPath());
EMAComponentInstanceSymbol instance = resolveComponentInstanceSymbol(qualifiedName, symtab);
generateFiles(symtab, instance, symtab, pythonPath, forced);
if (doCompile) {
compile();
}
processedArchitecture = null;
}
private EMAComponentInstanceSymbol resolveComponentInstanceSymbol(String qualifiedName, TaggingResolver symtab) {
EMAComponentSymbol component = symtab.<EMAComponentSymbol>resolve(qualifiedName, EMAComponentSymbol.KIND).orElse(null);
List<String> splitName = Splitters.DOT.splitToList(qualifiedName);
......@@ -126,15 +140,7 @@ public class EMADLGenerator {
System.exit(1);
}
EMAComponentInstanceSymbol instance = component.getEnclosingScope().<EMAComponentInstanceSymbol>resolve(instanceName, EMAComponentInstanceSymbol.KIND).get();
generateFiles(symtab, instance, symtab, pythonPath, forced);
if (doCompile) {
compile();
}
processedArchitecture = null;
return component.getEnclosingScope().<EMAComponentInstanceSymbol>resolve(instanceName, EMAComponentInstanceSymbol.KIND).get();
}
public void compile() throws IOException {
......@@ -530,7 +536,32 @@ public class EMADLGenerator {
final String fullConfigName = String.join(".", names);
ArchitectureSymbol correspondingArchitecture = this.processedArchitecture.get(fullConfigName);
assert correspondingArchitecture != null : "No architecture found for train " + fullConfigName + " configuration!";
configuration.setTrainedArchitecture(new ArchitectureAdapter(correspondingArchitecture));
configuration.setTrainedArchitecture(
new ArchitectureAdapter(correspondingArchitecture.getName(), correspondingArchitecture));
CNNTrainCocos.checkTrainedArchitectureCoCos(configuration);
// Resolve critic network if critic is present
if (configuration.getCriticName().isPresent()) {
String fullCriticName = configuration.getCriticName().get();
int indexOfFirstNameCharacter = fullCriticName.lastIndexOf('.') + 1;
fullCriticName = fullCriticName.substring(0, indexOfFirstNameCharacter)
+ fullCriticName.substring(indexOfFirstNameCharacter, indexOfFirstNameCharacter + 1).toUpperCase()
+ fullCriticName.substring(indexOfFirstNameCharacter + 1);
TaggingResolver symtab = EMADLAbstractSymtab.createSymTabAndTaggingResolver(getModelsPath());
EMAComponentInstanceSymbol instanceSymbol = resolveComponentInstanceSymbol(fullCriticName, symtab);
EMADLCocos.checkAll(instanceSymbol);
Optional<ArchitectureSymbol> critic = instanceSymbol.getSpannedScope().resolve("", ArchitectureSymbol.KIND);
if (!critic.isPresent()) {
Log.error("During the resolving of critic component: Critic component "
+ fullCriticName + " does not have a CNN implementation but is required to have one");
System.exit(-1);
}
critic.get().setComponentName(fullCriticName);
configuration.setCriticNetwork(new ArchitectureAdapter(fullCriticName, critic.get()));
CNNTrainCocos.checkCriticCocos(configuration);
}
cnnTrainGenerator.setInstanceName(componentInstance.getFullName().replaceAll("\\.", "_"));
Map<String, String> fileContentMap = cnnTrainGenerator.generateStrings(configuration);
......
package de.monticore.lang.monticar.emadl.generator;
package de.monticore.lang.monticar.emadl.generator.reinforcementlearning;
import de.monticore.lang.embeddedmontiarc.embeddedmontiarc._symboltable.instanceStructure.EMAComponentInstanceSymbol;
import de.monticore.lang.monticar.cnnarch.gluongenerator.reinforcement.RewardFunctionSourceGenerator;
import de.monticore.lang.monticar.emadl.generator.EMADLAbstractSymtab;
import de.monticore.lang.monticar.generator.cpp.GeneratorEMAMOpt2CPP;
import de.monticore.lang.tagging._symboltable.TaggingResolver;
import de.se_rwth.commons.logging.Log;
......@@ -9,30 +10,49 @@ import de.se_rwth.commons.logging.Log;
import java.io.IOException;
import java.util.Optional;
public class RewardFunctionCppGenerator implements RewardFunctionSourceGenerator {
public class RewardFunctionCppGenerator implements RewardFunctionSourceGenerator{
public RewardFunctionCppGenerator() {
}
@Override
public void generate(String modelPath, String rootModel, String targetPath) {
GeneratorEMAMOpt2CPP generator = new GeneratorEMAMOpt2CPP();
generator.useArmadilloBackend();
TaggingResolver taggingResolver = EMADLAbstractSymtab.createSymTabAndTaggingResolver(modelPath);
@Override
public EMAComponentInstanceSymbol resolveSymbol(TaggingResolver taggingResolver, String rootModel) {
Optional<EMAComponentInstanceSymbol> instanceSymbol = taggingResolver
.<EMAComponentInstanceSymbol>resolve(rootModel, EMAComponentInstanceSymbol.KIND);
if (!instanceSymbol.isPresent()) {
Log.error("Generation of reward function is not possible: Cannot resolve component instance "
+ rootModel);
+ rootModel);
}
return instanceSymbol.get();
}
@Override
public void generate(EMAComponentInstanceSymbol componentInstanceSymbol, TaggingResolver taggingResolver,
String targetPath) {
GeneratorEMAMOpt2CPP generator = new GeneratorEMAMOpt2CPP();
generator.useArmadilloBackend();
generator.setGenerationTargetPath(targetPath);
try {
generator.generate(instanceSymbol.get(), taggingResolver);
generator.generate(componentInstanceSymbol, taggingResolver);
} catch (IOException e) {
Log.error("Generation of reward function is not possible: " + e.getMessage());
}
}
@Override
public void generate(String modelPath, String rootModel, String targetPath) {
TaggingResolver taggingResolver = createTaggingResolver(modelPath);
EMAComponentInstanceSymbol instanceSymbol = resolveSymbol(taggingResolver, rootModel);
generate(instanceSymbol, taggingResolver, targetPath);
}
@Override
public TaggingResolver createTaggingResolver(final String modelPath) {
return EMADLAbstractSymtab.createSymTabAndTaggingResolver(modelPath);
}
}
......@@ -275,8 +275,8 @@ public class GenerationTest extends AbstractSymtabTest {
"HelperA.h",
"start_training.sh",
"reinforcement_learning/__init__.py",
"reinforcement_learning/CNNCreator_MountaincarCritic.py",
"reinforcement_learning/CNNNet_MountaincarCritic.py",
"reinforcement_learning/CNNCreator_mountaincar_agent_mountaincarCritic.py",
"reinforcement_learning/CNNNet_mountaincar_agent_mountaincarCritic.py",
"reinforcement_learning/strategy.py",
"reinforcement_learning/agent.py",
"reinforcement_learning/environment.py",
......
......@@ -17,7 +17,7 @@ configuration CartPoleDQN {
use_double_dqn : false
loss : euclidean
loss : huber
replay_memory : buffer{
memory_size : 10000
......
implementation Critic(state, action) {
(state ->
FullyConnected(units=400) ->
Relu() ->
FullyConnected(units=300)
|
action ->
FullyConnected(units=300)
) ->
Add() ->
Relu();
}
\ No newline at end of file
package mountaincar.agent;
component MountaincarCritic {
ports
in Q^{2} state,
in Q(-1:1)^{1} action,
out Q(-oo:oo)^{1} qvalues;
implementation CNN {
(
state ->
FullyConnected(units=400) ->
Relu() ->
FullyConnected(units=300)
|
action ->
FullyConnected(units=300)
) ->
Add() ->
Relu() ->
FullyConnected(units=1) ->
qvalues;
}
}
\ No newline at end of file
......@@ -23,7 +23,7 @@ configuration TorcsDQN {
use_double_dqn : true
loss : euclidean
loss : huber
replay_memory : buffer{
memory_size : 1000000
......
......@@ -3,8 +3,9 @@ import h5py
import mxnet as mx
import logging
import sys
from mxnet import nd
class cartpole_master_dqnDataLoader:
class CNNDataLoader_cartpole_master_dqn:
_input_names_ = ['state']
_output_names_ = ['qvalues_label']
......@@ -14,21 +15,38 @@ class cartpole_master_dqnDataLoader:
def load_data(self, batch_size):
train_h5, test_h5 = self.load_h5_files()
data_mean = train_h5[self._input_names_[0]][:].mean(axis=0)
data_std = train_h5[self._input_names_[0]][:].std(axis=0) + 1e-5
train_data = {}
data_mean = {}
data_std = {}
for input_name in self._input_names_:
train_data[input_name] = train_h5[input_name]
data_mean[input_name] = nd.array(train_h5[input_name][:].mean(axis=0))
data_std[input_name] = nd.array(train_h5[input_name][:].std(axis=0) + 1e-5)
train_label = {}
for output_name in self._output_names_:
train_label[output_name] = train_h5[output_name]
train_iter = mx.io.NDArrayIter(data=train_data,
label=train_label,
batch_size=batch_size)
train_iter = mx.io.NDArrayIter(train_h5[self._input_names_[0]],
train_h5[self._output_names_[0]],
batch_size=batch_size,
data_name=self._input_names_[0],
label_name=self._output_names_[0])
test_iter = None
if test_h5 != None:
test_iter = mx.io.NDArrayIter(test_h5[self._input_names_[0]],
test_h5[self._output_names_[0]],
batch_size=batch_size,
data_name=self._input_names_[0],
label_name=self._output_names_[0])
test_data = {}
for input_name in self._input_names_:
test_data[input_name] = test_h5[input_name]
test_label = {}
for output_name in self._output_names_:
test_label[output_name] = test_h5[output_name]
test_iter = mx.io.NDArrayIter(data=test_data,
label=test_label,
batch_size=batch_size)
return train_iter, test_iter, data_mean, data_std
def load_h5_files(self):
......@@ -36,21 +54,39 @@ class cartpole_master_dqnDataLoader:
test_h5 = None
train_path = self._data_dir + "train.h5"
test_path = self._data_dir + "test.h5"
if os.path.isfile(train_path):
train_h5 = h5py.File(train_path, 'r')
if not (self._input_names_[0] in train_h5 and self._output_names_[0] in train_h5):
logging.error("The HDF5 file '" + os.path.abspath(train_path) + "' has to contain the datasets: "
+ "'" + self._input_names_[0] + "', '" + self._output_names_[0] + "'")
sys.exit(1)
test_iter = None
for input_name in self._input_names_:
if not input_name in train_h5:
logging.error("The HDF5 file '" + os.path.abspath(train_path) + "' has to contain the dataset "
+ "'" + input_name + "'")
sys.exit(1)
for output_name in self._output_names_:
if not output_name in train_h5:
logging.error("The HDF5 file '" + os.path.abspath(train_path) + "' has to contain the dataset "
+ "'" + output_name + "'")
sys.exit(1)
if os.path.isfile(test_path):
test_h5 = h5py.File(test_path, 'r')
if not (self._input_names_[0] in test_h5 and self._output_names_[0] in test_h5):
logging.error("The HDF5 file '" + os.path.abspath(test_path) + "' has to contain the datasets: "
+ "'" + self._input_names_[0] + "', '" + self._output_names_[0] + "'")
sys.exit(1)
for input_name in self._input_names_:
if not input_name in test_h5:
logging.error("The HDF5 file '" + os.path.abspath(test_path) + "' has to contain the dataset "
+ "'" + input_name + "'")
sys.exit(1)
for output_name in self._output_names_:
if not output_name in test_h5:
logging.error("The HDF5 file '" + os.path.abspath(test_path) + "' has to contain the dataset "
+ "'" + output_name + "'")
sys.exit(1)
else:
logging.warning("Couldn't load test set. File '" + os.path.abspath(test_path) + "' does not exist.")
return train_h5, test_h5
else:
logging.error("Data loading failure. File '" + os.path.abspath(train_path) + "' does not exist.")
......
......@@ -101,7 +101,6 @@ class Net_0(gluon.HybridBlock):
self.fc3_ = gluon.nn.Dense(units=2, use_bias=True)
# fc3_, output shape: {[2,1,1]}
self.last_layers['qvalues'] = 'linear'
def hybrid_forward(self, F, state):
......
......@@ -56,7 +56,7 @@ if __name__ == "__main__":
'memory_size': 10000,
'sample_size': 32,
'state_dtype': 'float32',
'action_dtype': 'float32',
'action_dtype': 'uint8',
'rewards_dtype': 'float32'
},
'strategy_params': {
......@@ -78,10 +78,10 @@ if __name__ == "__main__":
'snapshot_interval': 20,
'max_episode_step': 250,
'target_score': 185.5,
'qnet':qnet_creator.net,
'qnet':qnet_creator.networks[0],
'use_fix_target': True,
'target_update_interval': 200,
'loss_function': 'euclidean',
'loss_function': 'huber',
'optimizer': 'rmsprop',
'optimizer_params': {
'learning_rate': 0.001 },
......@@ -108,4 +108,4 @@ if __name__ == "__main__":
train_successful = agent.train()
if train_successful:
agent.save_best_network(qnet_creator._model_dir_ + qnet_creator._model_prefix_ + '_0_newest', epoch=0)
agent.export_best_network(path=qnet_creator._model_dir_ + qnet_creator._model_prefix_ + '_0_newest', epoch=0)
......@@ -114,6 +114,8 @@ class Agent(object):
agent_session_file = os.path.join(session_dir, 'agent.p')
logger = self._logger
self._training_stats.save_stats(self._output_directory, episode=self._current_episode)
self._make_pickle_ready(session_dir)
with open(agent_session_file, 'wb') as f:
......@@ -122,10 +124,10 @@ class Agent(object):
def _make_pickle_ready(self, session_dir):
del self._training_stats.logger
self._logger = None
self._environment.close()
self._environment = None
self._save_net(self._best_net, 'best_net', session_dir)
self._export_net(self._best_net, 'best_net', filedir=session_dir)
self._logger = None
self._best_net = None
def _make_config_dict(self):
......@@ -177,6 +179,9 @@ class Agent(object):
return states, actions, rewards, next_states, terminals
def evaluate(self, target=None, sample_games=100, verbose=True):
if sample_games <= 0:
return 0
target = self._target_score if target is None else target
if target:
target_achieved = 0
......@@ -253,25 +258,22 @@ class Agent(object):
return self._target_score is not None\
and avg_reward > self._target_score
def _save_parameters(self, net, episode=None, filename='dqn-agent-params'):
def _export_net(self, net, filename, filedir=None, episode=None):
assert self._output_directory
assert isinstance(net, gluon.HybridBlock)
make_directory_if_not_exist(self._output_directory)
filedir = self._output_directory if filedir is None else filedir
filename = os.path.join(filedir, filename)
if(episode is not None):
self._logger.info(
'Saving model parameters after episode %d' % episode)
if episode is not None:
filename = filename + '-ep{}'.format(episode)
else:
self._logger.info('Saving model parameters')
self._save_net(net, filename)
def _save_net(self, net, filename, filedir=None):
filedir = self._output_directory if filedir is None else filedir
filename = os.path.join(filedir, filename + '.params')
net.save_parameters(filename)
net.export(filename, epoch=0)
net.save_parameters(filename + '.params')
def save_best_network(self, path, epoch=0):
def export_best_network(self, path=None, epoch=0):
path = os.path.join(self._output_directory, 'best_network')\
if path is None else path
self._logger.info(
'Saving best network with average reward of {}'.format(
self._best_avg_score))
......@@ -367,13 +369,17 @@ class DdpgAgent(Agent):
def _make_pickle_ready(self, session_dir):
super(DdpgAgent, self)._make_pickle_ready(session_dir)
self._save_net(self._actor, 'actor', session_dir)
self._export_net(self._actor, 'current_actor')
self._export_net(self._actor, 'actor', filedir=session_dir)
self._actor = None
self._save_net(self._critic, 'critic', session_dir)
self._export_net(self._critic, 'critic', filedir=session_dir)
self._critic = None
self._save_net(self._actor_target, 'actor_target', session_dir)
self._export_net(
self._actor_target, 'actor_target', filedir=session_dir)
self._actor_target = None
self._save_net(self._critic_target, 'critic_target', session_dir)
self._export_net(
self._critic_target, 'critic_target', filedir=session_dir)
self._critic_target = None
@classmethod
......@@ -441,7 +447,8 @@ class DdpgAgent(Agent):
return action[0].asnumpy()
def save_parameters(self, episode):
self._save_parameters(self._actor, episode=episode)
self._export_net(
self._actor, self._agent_name + '_actor', episode=episode)
def train(self, episodes=None):
self.save_config_file()
......@@ -457,9 +464,9 @@ class DdpgAgent(Agent):
else:
self._training_stats = DdpgTrainingStats(episodes)
# Initialize target Q' and mu'
self._actor_target = self._copy_actor()
self._critic_target = self._copy_critic()
# Initialize target Q' and mu'
self._actor_target = self._copy_actor()
self._critic_target = self._copy_critic()
# Initialize l2 loss for critic network
l2_loss = gluon.loss.L2Loss()
......@@ -496,6 +503,7 @@ class DdpgAgent(Agent):
# actor and exploration noise N according to strategy
action = self._strategy.select_action(
self.get_next_action(state))
self._strategy.decay(self._current_episode)
# Execute action a and observe reward r and next state ns
next_state, reward, terminal, _ = \
......@@ -537,10 +545,10 @@ class DdpgAgent(Agent):
# with critic parameters
tmp_critic = self._copy_critic()
with autograd.record():
actor_qvalues = tmp_critic(states, self._actor(states))
# For maximizing qvalues we have to multiply with -1
# as we use a minimizer
actor_loss = -1 * actor_qvalues
actor_loss = -tmp_critic(
states, self._actor(states)).mean()
actor_loss.backward()
trainer_actor.step(self._minibatch_size)
......@@ -558,7 +566,7 @@ class DdpgAgent(Agent):
episode_actor_loss +=\
np.sum(actor_loss.asnumpy()) / self._minibatch_size
episode_avg_q_value +=\
np.sum(actor_qvalues.asnumpy()) / self._minibatch_size
np.sum(qvalues.asnumpy()) / self._minibatch_size
training_steps += 1
......@@ -586,7 +594,6 @@ class DdpgAgent(Agent):
self._strategy.cur_eps, episode_reward)
self._do_snapshot_if_in_interval(self._current_episode)
self._strategy.decay(self._current_episode)
if self._is_target_reached(avg_reward):
self._logger.info(
......@@ -597,7 +604,7 @@ class DdpgAgent(Agent):
self._evaluate()
self.save_parameters(episode=self._current_episode)
self.save_best_network(os.path.join(self._output_directory, 'best'))
self.export_best_network()
self._training_stats.save_stats(self._output_directory)
self._logger.info('--------- Training finished ---------')
return True
......@@ -633,6 +640,359 @@ class DdpgAgent(Agent):
self._critic, self._state_dim, self._action_dim, ctx=self._ctx)
class TwinDelayedDdpgAgent(DdpgAgent):
def __init__(
self,
actor,
critic,
environment,
replay_memory_params,
strategy_params,
state_dim,
action_dim,
soft_target_update_rate=.001,
actor_optimizer='adam',
actor_optimizer_params={'learning_rate': 0.0001},
critic_optimizer='adam',
critic_optimizer_params={'learning_rate': 0.001},
ctx=None,
discount_factor=.9,
training_episodes=50,
start_training=20,
train_interval=1,
snapshot_interval=200,
agent_name='DdpgAgent',
max_episode_step=9999,
evaluation_samples=100,
output_directory='model_parameters',
verbose=True,
target_score=None,
policy_noise=0.2,
noise_clip=0.5,
policy_delay=2
):
super(TwinDelayedDdpgAgent, self).__init__(
environment=environment, replay_memory_params=replay_memory_params,
strategy_params=strategy_params, state_dim=state_dim,
action_dim=action_dim, ctx=ctx, discount_factor=discount_factor,
training_episodes=training_episodes, start_training=start_training,
train_interval=train_interval,
snapshot_interval=snapshot_interval, agent_name=agent_name,
max_episode_step=max_episode_step,
output_directory=output_directory, verbose=verbose,
target_score=target_score, evaluation_samples=evaluation_samples,
critic=critic, soft_target_update_rate=soft_target_update_rate,
actor=actor, actor_optimizer=actor_optimizer,
actor_optimizer_params=actor_optimizer_params,
critic_optimizer=critic_optimizer,
critic_optimizer_params=critic_optimizer_params)
self._policy_noise = policy_noise
self._noise_clip = noise_clip
self._policy_delay = policy_delay
self._critic2 = self._critic.__class__()
self._critic2.collect_params().initialize(
mx.init.Normal(), ctx=self._ctx)
self._critic2.hybridize()
self._critic2(nd.ones((1,) + state_dim, ctx=self._ctx),
nd.ones((1,) + action_dim, ctx=self._ctx))
self._critic2_target = self._copy_critic2()
self._critic2_optimizer = critic_optimizer
self._critic2_optimizer_params = self._adjust_optimizer_params(
critic_optimizer_params)
def _make_pickle_ready(self, session_dir):
super(TwinDelayedDdpgAgent, self)._make_pickle_ready(session_dir)
self._export_net(self._critic2, 'critic2', filedir=session_dir)
self._critic2 = None
self._export_net(
self._critic2_target, 'critic2_target', filedir=session_dir)
self._critic2_target = None
@classmethod
def resume_from_session(cls, session_dir, actor, critic, environment):
import pickle
if not os.path.exists(session_dir):
raise ValueError('Session directory does not exist')
files = dict()
files['agent'] = os.path.join(session_dir, 'agent.p')
files['best_net_params'] = os.path.join(session_dir, 'best_net.params')
files['actor_net_params'] = os.path.join(session_dir, 'actor.params')
files['actor_target_net_params'] = os.path.join(
session_dir, 'actor_target.params')
files['critic_net_params'] = os.path.join(session_dir, 'critic.params')
files['critic_target_net_params'] = os.path.join(
session_dir, 'critic_target.params')
files['critic2_net_params'] = os.path.join(
session_dir, 'critic2.params')
files['critic2_target_net_params'] = os.path.join(
session_dir, 'critic2_target.params')
for file in files.values():
if not os.path.exists(file):
raise ValueError(
'Session directory is not complete: {} is missing'